Baldy Dataset

A synthetic paired dataset of hair/bald image pairs for bald reconstruction research, generated from physically modeled hairstyles rendered on SMPL-X bodies in Blender.

Dataset Description

The Baldy dataset provides pixel-aligned, identity-consistent pairs of images showing the same person with and without hair. It was created using a multi-stage pipeline:

1. 3D Hair Modeling: Hairstyles from DiffLocks, Hair20K, USC-HairSalon, and CT2Hair are aligned to SMPL-X head meshes with varying poses, expressions, and optional BEDLAM clothing.
2. Blender Rendering: Hair is rendered under diverse camera views, lighting conditions, and material settings at 1024×1024 resolution.
3. Photorealistic Generation: ControlNet++ and SDXL-based pipelines generate photorealistic composites, with Flux Kontext refinement for identity preservation.

Dataset Statistics

Total samples: 6400

Samples per View

View	Count
back	99
front	6009
side	292

Samples per Hairstyle Source

Source	Count
ct2hair	9
difflocks	3197
hair20k	2824
usc	370

Data Fields

Field	Type	Description
hairstyle_id	string	Unique zero-padded sequential ID (e.g., "000042")
view	string	Camera view: "front", "back", or "side"
source	string	Hairstyle source: "difflocks", "hair20k", "usc", or "ct2hair"
hairstyle_source_id	string	Source-specific hairstyle identifier
hair_image	Image	Photorealistic image of the person with hair (1024×1024 PNG, auto-loaded)
bald_image	Image	Photorealistic image of the same person without hair (1024×1024 PNG, auto-loaded)
hair_render	Image	Blender-rendered hair on transparent background (1024×1024 PNG, auto-loaded)
background_image	Image	Scene background image (1024×1024 PNG, auto-loaded)
render_params_json	string	Full Blender render parameters as embedded JSON string
background_prompt	string	Text prompt used to generate the background (empty string if unavailable)
camera_focal_length	float	Camera focal length (mm)
camera_location_x	float	Camera X position in Blender world space
camera_location_y	float	Camera Y position in Blender world space
camera_location_z	float	Camera Z position in Blender world space
camera_rotation_x	float	Camera X rotation in radians
camera_rotation_y	float	Camera Y rotation in radians
camera_rotation_z	float	Camera Z rotation in radians
lighting_preset	string	Lighting preset name (e.g., "studio", "outdoor")
body_gender	string	Subject gender from SMPL-X body configuration
face_expression	string	Facial expression (e.g., "angry", "neutral"; empty string for some DiffLocks samples)
hair_melanin	float	Hair melanin value controlling color darkness
hair_roughness	float	Hair surface roughness
has_garments	bool	Whether BEDLAM clothing is applied to the body
views_available	string	All camera views available for this hairstyle (pipe-separated, e.g., "front|back|side")

Usage

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("deepmancer/baldy", split="train")

# Images are automatically decoded as PIL Images
sample = ds[0]
sample["hair_image"]        # PIL.Image (person with hair, auto-loaded)
sample["bald_image"]        # PIL.Image (person without hair, auto-loaded)
sample["hair_render"]       # PIL.Image (Blender hair render, auto-loaded)
sample["background_image"]  # PIL.Image (scene background, auto-loaded)

# Filter by view
front = ds.filter(lambda x: x["view"] == "front")
back  = ds.filter(lambda x: x["view"] == "back")
side  = ds.filter(lambda x: x["view"] == "side")

# Filter by hairstyle source
hair20k = ds.filter(lambda x: x["source"] == "hair20k")

# Access render parameters (embedded JSON)
import json
render_params = json.loads(sample["render_params_json"])

# Stream the dataset (no full download needed)
ds_stream = load_dataset("deepmancer/baldy", split="train", streaming=True)
for sample in ds_stream:
    print(sample["hairstyle_id"])
    break

File Structure

The dataset is stored as sharded Parquet files with embedded images (~1GB per shard):

data/
├── train-00000-of-NNNNN.parquet
├── train-00001-of-NNNNN.parquet
├── ...
└── train-NNNNN-of-NNNNN.parquet

Each Parquet file contains all columns including image bytes — no external files are needed. Images are decoded as PIL Images automatically when accessing rows.

Citation

If you use this dataset, please cite:

@article{baldy2025,
  title={HairPort: In-context 3D-Aware Hair Import and Transfer for Images},
  year={2025}
}

License

This dataset is released under the CC BY 4.0 license.